[Imageworld] Fast Deep Neural Nets Excel in Many Visual Pattern
juergen at idsia.ch
Fri Mar 9 11:57:07 CET 2012
A special breed of fast deep neural networks keeps winning important
pattern recognition competitions, lately even with human-competitive
New greatly improved world records on visual pattern recognition
• 3. NORB Object Recognition Benchmark. New record on full NORB: 2.7%
error rate. The best result by others (brbo) is 5%.
• 2. CIFAR-10 Object Recognition Benchmark. New record 11.2% (brbo
• 1. MNIST Handwritten Digit Recognition (perhaps the most famous
machine learning benchmark). New record 0.23%. This is the first human-
competitive result on MNIST (brbo 0.39%).
Upcoming paper on this with Dan Cireşan & Ueli Meier at CVPR 2012:
Multi-Column Deep Neural Networks for Image Classification
(long preprint available)
1st ranks in visual pattern recognition competitions:
• 7. NEW: March 2012: ISBI 2012 Segmentation Challenge, won on all
three evaluation metrics by a large margin, with superhuman pixel
error rate (with Dan Cireşan & Alessandro Giusti)
• 6. August 2011: IJCNN 2011 on-site Traffic Sign Recognition
Competition: 0.56% error rate - nearly three times better than brbo.
The only method outperforming humans.
• 5. June 2011: ICDAR 2011 offline Chinese Handwriting Recognition
• 4. January 2011: Online German Traffic Sign Recognition Contest
(1st & 2nd rank)
• 3. ICDAR 2009 Arabic Connected Handwriting Competition, like the
others below won by LSTM recurrent neural networks (deep by nature).
• 2. ICDAR 2009 Handwritten Farsi/Arabic Character Recognition
• 1. ICDAR 2009 French Connected Handwriting Competition
No unsupervised pre-training.
Overview web site with more details and Google Tech Talk and
contributions to neural net vision applications by authors including
Dan Cireşan, Ueli Meier, Jonathan Masci, Alessandro Giusti, Alex
Graves, Jawad Nagi, Frederic Ducatelle, Gianni Di Caro, Luca Maria
Director of the Swiss AI Lab IDSIA, Lugano
Professor of Artificial Intelligence, Univ. Lugano
Professor SUPSI, Manno-Lugano, Switzerland
More information about the Imageworld